Graphcore is the most instructive cautionary tale in AI hardware and also the source of some genuinely good architectural teaching, which makes reviewing its learning material an odd exercise. The company built the Intelligence Processing Unit around a thesis that was intellectually serious: that the memory bandwidth bottleneck in GPU training could be addressed by putting a very large amount of SRAM directly on the chip, close to a very large number of small independent cores, and running a bulk synchronous parallel execution model where compute and exchange phases alternate in lockstep. That is a different machine from a GPU, not a variation on one, and understanding why someone would build it teaches you a great deal about where the actual constraints in deep learning workloads are. The Poplar SDK is the software layer that makes the architecture usable, compiling machine learning graphs from PyTorch and TensorFlow down to the IPU's tile structure and providing the libraries and profiling tools you develop against.
The tutorial repository on GitHub covers tutorials, feature examples and simple applications, and the material is hands-on and written by people who understood the machine. The profiling material in particular is worth reading even if you never run it, because IPU profiling forces you to think about the compute and exchange phases explicitly in a way that GPU profiling generally does not. Reading it changes how you think about data movement costs, which transfers. The problem is everything around the software.
Graphcore raised at a 2.8 billion dollar valuation, was talked about seriously as a European challenger to Nvidia, failed to convert that into meaningful commercial traction, and was acquired by SoftBank. It is now funded, with a reported 457 million dollar injection in 2026, a new engineering campus in Bengaluru, and a role in SoftBank's broader infrastructure programme including co-development with Ampere on the Izanagi chip aimed at hyperscale data centre deployment, with the Poplar SDK merging into a unified toolchain across both architectures. That is survival, and it is a fundamentally different proposition from what the developer materials were written for. A company pursuing broad developer adoption publishes tutorials, maintains cloud access so anyone can try the hardware, and cultivates an external ecosystem.
A company co-developing silicon for a small number of enormous data centre deployments does not need any of that, and the public developer surface has contracted accordingly. Cloud IPU access, which was once available through several providers and made the tutorials genuinely runnable, is now difficult to obtain. Much of the published material has not kept pace. You can still read it, and you increasingly cannot run it.
My honest recommendation splits the material in two. The architectural and conceptual content is worth a few hours to anyone interested in AI hardware. Understanding why on-chip memory changes the performance calculus, what bulk synchronous parallelism buys and costs, and why a machine optimised for fine-grained parallelism handles some model shapes very differently from a GPU, is education that applies when you read about any accelerator, including the ones you will actually use. The hands-on Poplar work is only worth committing to if you have hardware access through an employer or a research programme.
As a career skill, IPU expertise has essentially no market. That is not a criticism of the engineering, which was real, and it is the situation. The wider lesson is one worth internalising before betting learning time on any challenger accelerator ecosystem: the software stack and the developer community are the product, the silicon is the easy part, and a company that has to choose between courting developers and serving one very large customer will choose the customer every time. Three point zero.
Genuinely good architectural teaching about a distinctive machine, published by a company that no longer has a reason to invest in public developer education, with hands-on material you will struggle to run.